# Lakebase Search Boosts Efficiency and Cost-Effectiveness in Hybrid Search

> Discover how Lakebase Search transforms application development by integrating hybrid search directly into databases, achieving low latency and cost-efficient operations for data-driven applications.

**Source**: databricks.com | **Published**: 2026-10-15 | **Type**: case_study

## Key Facts

- Hybrid search in Lakebase Postgres keeps 90.5% of top results with 200 candidates, enhancing efficiency.
- Cost-effective compute at $0.02 for 109k requests shows potential for scalable, low-cost operations.
- Cache size matching vector index reduces latency significantly, indicating optimal resource allocation.
- Cold starts impact user experience; maintaining 0.5 CU minimizes delays, crucial for user retention.
- Scaling app workers beyond 8 CU maximizes throughput, revealing strategic focus on application performance.

## Summary

\## Summary
Databricks deployed Lakebase Search within its internal Slide Hub application to enhance search capabilities for field engineers accessing a repository of 700 presentations. The challenge was to improve search efficiency and reduce latency without maintaining a separate search system. By integrating Lakebase Search, the application achieved a significant reduction in search latency and improved throughput, demonstrating the effectiveness of hybrid search methods.

\## Background
Databricks, a leader in data and AI solutions, developed Slide Hub as an internal application for its field engineers. The app allows users to find and reuse slides from approximately 700 internal presentations, including event talks and product decks. Prior to the deployment of Lakebase Search, the search functionality was limited, leading to inefficiencies in retrieving relevant slides quickly.

\## Challenge
The primary challenge was to enhance the search functionality within Slide Hub, specifically to reduce latency and improve the relevance of search results. The existing system required a more efficient method to handle search queries without the overhead of a separate search infrastructure.

\## Solution
Lakebase Search was implemented as a set of Postgres extensions directly within the Slide Hub application database. This integration allowed for hybrid search capabilities, combining vector search and BM25 keyword search. The system was designed to maintain a single transaction path, eliminating the need for data duplication. The application utilized a Local File Cache to optimize search performance, ensuring that a majority of pages remained in memory for quick access.

\## Results
The deployment of Lakebase Search resulted in substantial performance improvements. For instance, with a compute unit (CU) configuration of 16 CU and a MEDIUM app size, the system achieved at least 662 requests per second with a median latency of 343 milliseconds. Additionally, the candidate limit was optimized to 200, maintaining 90.5% of the top-10 results while significantly reducing the number of pages read during searches.

\## Key Insights
1. **Hybrid Search Efficiency**: Combining vector and keyword search methods can yield better results and faster response times.
2. **Cache Optimization**: Sizing the Local File Cache to match the vector index can drastically reduce latency and improve hit rates.
3. **Scalability Considerations**: Understanding the limits of both the database and application compute resources is crucial for optimizing performance.

\## Customer Testimonial
No direct quotes were provided in the source material.

## Entities

- **Companies**: Databricks
- **Products**: Lakebase Search, Lakebase Postgres, Slide Hub
- **Technologies**: Postgres, BM25, vector search, ANN, SQL

## Key Concepts

hybrid search, Local File Cache, compute units, candidate pools, cache ratio, recall, latency, cost management

## Definitions

- **hybrid search**: A search method that combines vector search and BM25 keyword search to find relevant results.
- **Local File Cache**: A memory cache that stores recently read pages to reduce latency in database queries.
- **compute units (CU)**: A measure of compute resources allocated to Lakebase Postgres, where one CU typically provides about 2 GB of RAM.
- **candidate pool**: The set of potential results returned by a search query, which can be limited to improve performance.
- **recall**: A metric that measures the proportion of relevant results returned by a search compared to the total number of relevant results.

## Use Cases

- Searching internal presentations in Slide Hub
- Benchmarking database performance under load
- Optimizing search latency with candidate limits
- Scaling compute resources based on query load
- Managing costs associated with database usage

## Frequently Asked Questions

**What is Lakebase Search?**

Lakebase Search is a set of Postgres extensions that allows for hybrid search capabilities directly within the application database, eliminating the need for a separate search system.

**How does the Local File Cache improve performance?**

The Local File Cache stores frequently accessed pages in memory, which significantly reduces the latency of search queries by minimizing the need to read from slower storage.

**What are compute units and how do they affect performance?**

Compute units (CU) represent the amount of computational resources allocated to Lakebase Postgres. More CUs provide additional memory and processing power, which can enhance query performance and throughput.

**Why is it important to limit the candidate pool?**

Limiting the candidate pool helps to maintain high performance by reducing the number of pages read during a query, which in turn lowers latency while still preserving a high percentage of relevant results.

**How can I manage costs when using Lakebase Postgres?**

To manage costs effectively, consider setting a minimum compute unit allocation to avoid cold starts, and scale down resources during periods of low usage, as Lakebase Postgres only bills for compute when it is actively running.

## Links

- [Read on Welcome.AI](https://welcome.ai/content/lakebase-search-boosts-efficiency-and-cost-effectiveness-in-hybrid-search)
- [Original source](https://www.databricks.com/blog/sizing-lakebase-search-lessons-real-app-and-4-million-passages)
- [Databricks](https://welcome.ai/company/databricks): Featured company

---

Source: Welcome.AI | https://welcome.ai/content/lakebase-search-boosts-efficiency-and-cost-effectiveness-in-hybrid-search